t/nvmept_trim: increase transfer size for some tests
[fio.git] / backend.c
index cb1fbf424136d03de69774d0b36112c5789857a7..fe03eab3870f1588bc5e09ab119e7d358d659f52 100644 (file)
--- a/backend.c
+++ b/backend.c
@@ -49,6 +49,7 @@
 #include "helper_thread.h"
 #include "pshared.h"
 #include "zone-dist.h"
+#include "fio_time.h"
 
 static struct fio_sem *startup_sem;
 static struct flist_head *cgroup_list;
@@ -93,19 +94,16 @@ static void sig_int(int sig)
 #ifdef WIN32
 static void sig_break(int sig)
 {
-       struct thread_data *td;
-       int i;
-
        sig_int(sig);
 
        /**
         * Windows terminates all job processes on SIGBREAK after the handler
         * returns, so give them time to wrap-up and give stats
         */
-       for_each_td(td, i) {
+       for_each_td(td) {
                while (td->runstate < TD_EXITED)
                        sleep(1);
-       }
+       } end_for_each();
 }
 #endif
 
@@ -469,7 +467,7 @@ int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
                                if (!from_verify)
                                        unlog_io_piece(td, io_u);
                                td_verror(td, EIO, "full resid");
-                               put_io_u(td, io_u);
+                               clear_io_u(td, io_u);
                                break;
                        }
 
@@ -637,15 +635,6 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
        if (td->error)
                return;
 
-       /*
-        * verify_state needs to be reset before verification
-        * proceeds so that expected random seeds match actual
-        * random seeds in headers. The main loop will reset
-        * all random number generators if randrepeat is set.
-        */
-       if (!td->o.rand_repeatable)
-               td_fill_verify_state_seed(td);
-
        td_set_runstate(td, TD_VERIFYING);
 
        io_u = NULL;
@@ -909,7 +898,16 @@ static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
        if (left)
                total = usec_spin(left);
 
-       left = td->o.thinktime - total;
+       /*
+        * usec_spin() might run for slightly longer than intended in a VM
+        * where the vCPU could get descheduled or the hypervisor could steal
+        * CPU time. Ensure "left" doesn't become negative.
+        */
+       if (total < td->o.thinktime)
+               left = td->o.thinktime - total;
+       else
+               left = 0;
+
        if (td->o.timeout) {
                runtime_left = td->o.timeout - utime_since_now(&td->epoch);
                if (runtime_left < (unsigned long long)left)
@@ -979,6 +977,11 @@ static void do_io(struct thread_data *td, uint64_t *bytes_done)
        */
        if (td_write(td) && td_random(td) && td->o.norandommap)
                total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
+
+       /* Don't break too early if io_size > size */
+       if (td_rw(td) && !td_random(td))
+               total_bytes = max(total_bytes, (uint64_t)td->o.io_size);
+
        /*
         * If verify_backlog is enabled, we'll run the verify in this
         * handler as well. For that case, we may need up to twice the
@@ -1049,8 +1052,11 @@ static void do_io(struct thread_data *td, uint64_t *bytes_done)
                }
 
                if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY) {
-                       io_u->numberio = td->io_issues[io_u->ddir];
-                       populate_verify_io_u(td, io_u);
+                       if (!(io_u->flags & IO_U_F_PATTERN_DONE)) {
+                               io_u_set(td, io_u, IO_U_F_PATTERN_DONE);
+                               io_u->numberio = td->io_issues[io_u->ddir];
+                               populate_verify_io_u(td, io_u);
+                       }
                }
 
                ddir = io_u->ddir;
@@ -1133,6 +1139,9 @@ reap:
                if (ret < 0)
                        break;
 
+               if (ddir_rw(ddir) && td->o.thinkcycles)
+                       cycles_spin(td->o.thinkcycles);
+
                if (ddir_rw(ddir) && td->o.thinktime)
                        handle_thinktime(td, ddir, &comp_time);
 
@@ -1329,7 +1338,7 @@ static int init_io_u(struct thread_data *td)
 int init_io_u_buffers(struct thread_data *td)
 {
        struct io_u *io_u;
-       unsigned long long max_bs, min_write;
+       unsigned long long max_bs, min_write, trim_bs = 0;
        int i, max_units;
        int data_xfer = 1;
        char *p;
@@ -1340,7 +1349,18 @@ int init_io_u_buffers(struct thread_data *td)
        td->orig_buffer_size = (unsigned long long) max_bs
                                        * (unsigned long long) max_units;
 
-       if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
+       if (td_trim(td) && td->o.num_range > 1) {
+               trim_bs = td->o.num_range * sizeof(struct trim_range);
+               td->orig_buffer_size = trim_bs
+                                       * (unsigned long long) max_units;
+       }
+
+       /*
+        * For reads, writes, and multi-range trim operations we need a
+        * data buffer
+        */
+       if (td_ioengine_flagged(td, FIO_NOIO) ||
+           !(td_read(td) || td_write(td) || (td_trim(td) && td->o.num_range > 1)))
                data_xfer = 0;
 
        /*
@@ -1392,7 +1412,10 @@ int init_io_u_buffers(struct thread_data *td)
                                fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
                        }
                }
-               p += max_bs;
+               if (td_trim(td) && td->o.num_range > 1)
+                       p += trim_bs;
+               else
+                       p += max_bs;
        }
 
        return 0;
@@ -1642,7 +1665,7 @@ static void *thread_main(void *data)
        uint64_t bytes_done[DDIR_RWDIR_CNT];
        int deadlock_loop_cnt;
        bool clear_state;
-       int res, ret;
+       int ret;
 
        sk_out_assign(sk_out);
        free(fd);
@@ -1799,13 +1822,16 @@ static void *thread_main(void *data)
 
        /* ioprio_set() has to be done before td_io_init() */
        if (fio_option_is_set(o, ioprio) ||
-           fio_option_is_set(o, ioprio_class)) {
-               ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
+           fio_option_is_set(o, ioprio_class) ||
+           fio_option_is_set(o, ioprio_hint)) {
+               ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class,
+                                o->ioprio, o->ioprio_hint);
                if (ret == -1) {
                        td_verror(td, errno, "ioprio_set");
                        goto err;
                }
-               td->ioprio = ioprio_value(o->ioprio_class, o->ioprio);
+               td->ioprio = ioprio_value(o->ioprio_class, o->ioprio,
+                                         o->ioprio_hint);
                td->ts.ioprio = td->ioprio;
        }
 
@@ -1855,7 +1881,7 @@ static void *thread_main(void *data)
        if (rate_submit_init(td, sk_out))
                goto err;
 
-       set_epoch_time(td, o->log_unix_epoch | o->log_alternate_epoch, o->log_alternate_epoch_clock_id);
+       set_epoch_time(td, o->log_alternate_epoch_clock_id, o->job_start_clock_id);
        fio_getrusage(&td->ru_start);
        memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
        memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
@@ -1894,8 +1920,12 @@ static void *thread_main(void *data)
                if (td->o.verify_only && td_write(td))
                        verify_bytes = do_dry_run(td);
                else {
+                       if (!td->o.rand_repeatable)
+                               /* save verify rand state to replay hdr seeds later at verify */
+                               frand_copy(&td->verify_state_last_do_io, &td->verify_state);
                        do_io(td, bytes_done);
-
+                       if (!td->o.rand_repeatable)
+                               frand_copy(&td->verify_state, &td->verify_state_last_do_io);
                        if (!ddir_rw_sum(bytes_done)) {
                                fio_mark_td_terminate(td);
                                verify_bytes = 0;
@@ -1979,13 +2009,23 @@ static void *thread_main(void *data)
         * another thread is checking its io_u's for overlap
         */
        if (td_offload_overlap(td)) {
-               int res = pthread_mutex_lock(&overlap_check);
-               assert(res == 0);
+               int res;
+
+               res = pthread_mutex_lock(&overlap_check);
+               if (res) {
+                       td->error = errno;
+                       goto err;
+               }
        }
        td_set_runstate(td, TD_FINISHING);
        if (td_offload_overlap(td)) {
+               int res;
+
                res = pthread_mutex_unlock(&overlap_check);
-               assert(res == 0);
+               if (res) {
+                       td->error = errno;
+                       goto err;
+               }
        }
 
        update_rusage_stat(td);
@@ -2058,16 +2098,15 @@ err:
 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
                         uint64_t *m_rate)
 {
-       struct thread_data *td;
        unsigned int cputhreads, realthreads, pending;
-       int i, status, ret;
+       int ret;
 
        /*
         * reap exited threads (TD_EXITED -> TD_REAPED)
         */
        realthreads = pending = cputhreads = 0;
-       for_each_td(td, i) {
-               int flags = 0;
+       for_each_td(td) {
+               int flags = 0, status;
 
                if (!strcmp(td->o.ioengine, "cpuio"))
                        cputhreads++;
@@ -2159,7 +2198,7 @@ reaped:
                done_secs += mtime_since_now(&td->epoch) / 1000;
                profile_td_exit(td);
                flow_exit_job(td);
-       }
+       } end_for_each();
 
        if (*nr_running == cputhreads && !pending && realthreads)
                fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
@@ -2286,13 +2325,11 @@ static bool waitee_running(struct thread_data *me)
 {
        const char *waitee = me->o.wait_for;
        const char *self = me->o.name;
-       struct thread_data *td;
-       int i;
 
        if (!waitee)
                return false;
 
-       for_each_td(td, i) {
+       for_each_td(td) {
                if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
                        continue;
 
@@ -2302,7 +2339,7 @@ static bool waitee_running(struct thread_data *me)
                                        runstate_to_name(td->runstate));
                        return true;
                }
-       }
+       } end_for_each();
 
        dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
        return false;
@@ -2326,14 +2363,14 @@ static void run_threads(struct sk_out *sk_out)
        set_sig_handlers();
 
        nr_thread = nr_process = 0;
-       for_each_td(td, i) {
+       for_each_td(td) {
                if (check_mount_writes(td))
                        return;
                if (td->o.use_thread)
                        nr_thread++;
                else
                        nr_process++;
-       }
+       } end_for_each();
 
        if (output_format & FIO_OUTPUT_NORMAL) {
                struct buf_output out;
@@ -2359,7 +2396,7 @@ static void run_threads(struct sk_out *sk_out)
        nr_started = 0;
        m_rate = t_rate = 0;
 
-       for_each_td(td, i) {
+       for_each_td(td) {
                print_status_init(td->thread_number - 1);
 
                if (!td->o.create_serialize)
@@ -2395,7 +2432,7 @@ reap:
                                        td_io_close_file(td, f);
                        }
                }
-       }
+       } end_for_each();
 
        /* start idle threads before io threads start to run */
        fio_idle_prof_start();
@@ -2411,7 +2448,7 @@ reap:
                /*
                 * create threads (TD_NOT_CREATED -> TD_CREATED)
                 */
-               for_each_td(td, i) {
+               for_each_td(td) {
                        if (td->runstate != TD_NOT_CREATED)
                                continue;
 
@@ -2490,7 +2527,7 @@ reap:
 
                                        ret = (int)(uintptr_t)thread_main(fd);
                                        _exit(ret);
-                               } else if (i == fio_debug_jobno)
+                               } else if (__td_index == fio_debug_jobno)
                                        *fio_debug_jobp = pid;
                                free(eo);
                                free(fd);
@@ -2506,7 +2543,7 @@ reap:
                                break;
                        }
                        dprint(FD_MUTEX, "done waiting on startup_sem\n");
-               }
+               } end_for_each();
 
                /*
                 * Wait for the started threads to transition to
@@ -2551,7 +2588,7 @@ reap:
                /*
                 * start created threads (TD_INITIALIZED -> TD_RUNNING).
                 */
-               for_each_td(td, i) {
+               for_each_td(td) {
                        if (td->runstate != TD_INITIALIZED)
                                continue;
 
@@ -2565,7 +2602,7 @@ reap:
                        t_rate += ddir_rw_sum(td->o.rate);
                        todo--;
                        fio_sem_up(td->sem);
-               }
+               } end_for_each();
 
                reap_threads(&nr_running, &t_rate, &m_rate);
 
@@ -2591,9 +2628,7 @@ static void free_disk_util(void)
 
 int fio_backend(struct sk_out *sk_out)
 {
-       struct thread_data *td;
        int i;
-
        if (exec_profile) {
                if (load_profile(exec_profile))
                        return 1;
@@ -2649,7 +2684,7 @@ int fio_backend(struct sk_out *sk_out)
                }
        }
 
-       for_each_td(td, i) {
+       for_each_td(td) {
                struct thread_stat *ts = &td->ts;
 
                free_clat_prio_stats(ts);
@@ -2662,7 +2697,7 @@ int fio_backend(struct sk_out *sk_out)
                }
                fio_sem_remove(td->sem);
                td->sem = NULL;
-       }
+       } end_for_each();
 
        free_disk_util();
        if (cgroup_list) {